agent-teams

name	agent-teams
version	1.0.0
description	Team composition templates for coordinated multi-agent tasks
author	iscmga
tags	["agents","teams","orchestration","roles"]
triggers	{"keywords":["team","swarm","multi-agent","coordinate","complex task","large task","feature","refactor"]}

Agent Teams

Principle

Match team composition to task type. Smaller, focused teams outperform large swarms. Each role has a clear responsibility and produces a defined artifact.

Team Templates

Bug Fix (4 agents)

Role	Responsibility	Produces
Researcher	Reproduce bug, trace root cause, gather context	handoffs/research.md — reproduction steps, root cause analysis
Implementer	Write the fix based on research findings	The code change
Tester	Write/update tests that cover the fix	Test files
Reviewer	Review fix for correctness, side effects, regressions	handoffs/review.md — approval or change requests

Sequence: Researcher → Implementer → Tester + Reviewer (parallel)

Feature Development (5 agents)

Role	Responsibility	Produces
Architect	Design approach, identify affected components	handoffs/design.md — approach, interfaces, risks
Implementer	Build the feature per design	The code
Tester	Write unit + integration tests	Test files
Reviewer	Code quality, security, patterns compliance	handoffs/review.md
Documenter	Update docs, README, inline comments if needed	Doc changes

Sequence: Architect → Implementer → Tester + Reviewer + Documenter (parallel)

Refactor (3 agents)

Role	Responsibility	Produces
Analyst	Map current code, identify dependencies, plan changes	handoffs/analysis.md — dependency map, risk areas
Implementer	Execute refactor per analysis	The code changes
Reviewer	Verify behavior preservation, no regressions	handoffs/review.md

Sequence: Analyst → Implementer → Reviewer

Security Audit (3 agents)

Role	Responsibility	Produces
Scanner	Run automated checks (Prowler, Security Hub, OWASP)	handoffs/scan-results.md
Analyst	Triage findings, assess severity, identify false positives	handoffs/triage.md — prioritized findings
Remediator	Fix critical/high findings	The code fixes

Sequence: Scanner → Analyst → Remediator

Performance Investigation (3 agents)

Role	Responsibility	Produces
Profiler	Identify bottlenecks, gather metrics	handoffs/profile.md — bottleneck analysis
Optimizer	Implement targeted optimizations	The code changes
Validator	Benchmark before/after, verify improvements	handoffs/benchmark.md

Sequence: Profiler → Optimizer → Validator

Infrastructure Change (4 agents)

Role	Responsibility	Produces
Architect	Design infrastructure change, assess blast radius	handoffs/design.md
Implementer	Write Terraform/IaC code	The .tf files
Security Reviewer	Check IAM, network, encryption, compliance	handoffs/security-review.md
Validator	Run plan, verify no unintended changes	handoffs/plan-output.md

Sequence: Architect → Implementer → Security Reviewer + Validator (parallel)

How to Use

Claude Code

The coordinator (your main session) follows this pattern:

1. Identify task type → pick the matching team template
2. Write coordinator.md with task decomposition (see agent-coordination skill)
3. Spawn agents sequentially or in parallel per the template's sequence using the Agent tool
4. Each agent reads coordinator.md, does its work, writes its handoff file
5. Coordinator reads all handoffs and integrates

Cursor

Run each role as a separate chat session, passing context through the handoff files in the workspace.

Guidelines

• Don't create teams for simple tasks — one agent is fine for small changes
• Use the smallest team that covers the task — 3 agents beats 6 if 3 is enough
• Every agent must have exactly one clear deliverable
• Parallel agents must not modify the same files
• The coordinator is always human-in-the-loop for destructive operations